EricJaebeom/CBSR-NIR-Face

CBSR NIR Face Dataset

This is a byte-preserving mirror of Gabriel Preda's CBSR NIR Face Dataset on Kaggle, version 1 (2024-11-18). The original dataset is attributed to CBSR. The description below is adapted from the Kaggle data card by Gabriel Preda; this mirror adds Hugging Face metadata, file layout, provenance, and clarification of eye-side conventions.

Data

All 3,942 source files are preserved at their original paths and with their original bytes. There are 3,940 uncompressed, 8-bit grayscale BMP images of 197 people. Each image has width 480 and height 640 pixels. All images are near-infrared (NIR).

The gallery contains 1,576 images (8 per person), and the probe contains 2,364 images (12 per person).

NIR_face_dataset/NIR_face_dataset/*.bmp
gallery-groundtruth.txt
probe-groundtruth.txt
source-manifest.json

Eye-coordinate annotations

The ground-truth files contain headerless CSV records:

filename,image_left_x,image_left_y,image_right_x,image_right_y

Filenames follow [person id]-[image id]-a[-g].bmp; the optional -g indicates glasses. Coordinates are expressed in original-image pixels.

For example, 0001-00-a-g.bmp,197,345,325,349 identifies person 0001 with glasses and image-side eye positions (197, 345) and (325, 349). 0007-22-a.bmp,175,182,285,179 identifies person 0007 without glasses.

Left and right in these source annotations refer to image sides, not anatomical eye sides. Image-left corresponds to anatomical right (OD); image-right corresponds to anatomical left (OS). The source files in this mirror are unchanged. Consumers using these native eye positions as iris-center labels must apply this semantic mapping.

Provenance and license

Kaggle lists this dataset under the MIT license; this card carries that upstream declaration. The mirror does not introduce a new license or claim authorship of the data. Please retain attribution to CBSR and to the Kaggle uploader, Gabriel Preda.

source-manifest.json records the Kaggle version, download archive SHA-256, and original path, byte size, and SHA-256 of every source file. README and the manifest are mirror metadata, not original dataset files. No images or annotations have been generated, transformed, or filtered.

Download

Use huggingface_hub.snapshot_download("EricJaebeom/CBSR-NIR-Face", repo_type="dataset"). No authentication is required. Pin the repository revision for reproducible use. The original ground-truth text files remain separate from the image files; this repository is a source-file mirror.